Delegate, don’t chat: Claude Tag moves into Slack
- Agentic Engineering
- Coding Agents
- Automation
- Claude
Anthropic has introduced Claude Tag, and the name nails what it is about. You tag @Claude in a Slack channel, give it a task in plain language, and the agent works it off. Not in a separate chat window, but where your team already talks. Anthropic calls it the beginning of an evolution of Claude Code. I find that more honest than „a bot for Slack“, because the difference is not the place, it is the way of working.
What is actually new
Three things lift Claude Tag above the familiar bot.
Shared context. In a channel everyone works with the same Claude instance, not each with their own conversation. Whoever joins later sees the ongoing work and picks it up, without having to re-explain the state. Claude remembers what is relevant in the channel.
Proactive. Turn on the ambient behavior and Claude speaks up on its own. It flags things you might need to know and follows up on threads that were left hanging.
Asynchronous. You hand over a task and turn to something else. Claude schedules steps for itself and pursues a project over hours or days. Under the hood runs Opus 4.8.
Anthropic backs this with a number from its own house: 65 percent of the product team’s code already comes through the internal version. And it does not stop at code, it moves into metrics, support tickets, and bug investigation. The launch is a beta for Enterprise and Team customers, it replaces the existing Claude in Slack app, and admins have 30 days to opt in.
From prompting to delegating
I recently wrote here about loop engineering: stop prompting agents, build loops that prompt them. Claude Tag is exactly that loop, only packaged and dropped into Slack. The cadence, the scheduling of its own tasks, the memory beyond the conversation, the proactive nudging. What you stitched together from bash scripts last year now ships as a product.
That is the real shift. You no longer chat with a tool, you delegate to a colleague who sits in the same channel. Sounds like a small thing. But it changes who makes the first move. When you chat, you start. With a proactive agent, sometimes it does.
Old problems get sharper, a new one appears
This is exactly where I stay skeptical, for the same reasons as always. Verification stays with you. An agent working unattended over days also makes mistakes unattended over days. „Done“ in the thread is a claim, not a proof.
And your understanding erodes if you let it. The more work appears in the channel that nobody wrote themselves, the wider the gap between what sits in the repo and what the team has actually grasped.
The shared channel brings a new problem the private chat did not have: responsibility diffuses. When everyone sees the same agent and it kicks off something on its own, soon nobody feels in charge of checking the result. Each assumes someone else is looking. That is not an AI problem, it is an old team problem. A proactive agent in a shared channel only amplifies it.
Anthropic built sensible levers for this. Admins decide per channel which tools and data Claude can reach, separate instances with their own memories, set token budgets per channel and organization, and see in the log which task came from whom. Sales data stays out of the engineering channel. That handles access and cost. It does not handle who, in the end, reads what the agent built.
Treat it like a new teammate
Try it, the step is real and it is big. But treat Claude Tag like a new teammate, not an oracle. A new teammate gets clear tasks at first, narrow access, and someone who reviews the work. That is exactly how I would start: a private channel, a tight mandate, a human who stays responsible.
The leverage point moves further, away from the single prompt, toward setting up and fencing in an agent that works alongside you. That does not make the work easier. It makes it different. Delegating is a leadership job, and leadership does not mean looking away.